Peter Lai and the business hidden in the nuisance

Peter Lai once looked at a business opportunity sitting inside his own startup and doubted there was a business there. The company was building tools for working with documents online. Customers wanted to put its document viewer inside their own products. Lai favored other directions, including electronic signatures. The customers had a different agenda. A useful piece of software was trying to escape the product that contained it.

It is a pleasingly inconvenient episode in a founder’s career. In the polished version of entrepreneurship, someone spots the opportunity and everyone catches up. Here, one of the builders needed persuading. Crocodoc’s document-viewing technology eventually went to Box in 2013. Lai went with it, then left to build the original Emburse corporate-card startup. His career rewards a closer look at the tasks between the milestones.

The document that would not behave

For anyone who worked online in the early 2010s, the irritation was familiar. A PDF or Office document arrived, and reading it could require a download, another application or a detour through awkward formatting. A photograph could sit comfortably in a browser. A document sometimes acted as though it needed its own travel arrangements.

Crocodoc converted documents into HTML so people could view, share and annotate them online. Its customers included LinkedIn, Yammer, Blackboard, Edmodo and SAP. The technical work happened underneath the thing a person wanted to do: open the file and get on with it. By October 2013, the company reported more than 200 million document conversions and 14 million annotations.

The founding team came through MIT. Ryan Damico, Bennett Rogers, Matt Long and Lai had worked on WebNotes, a research annotation product. Crocodoc emerged from that earlier work. Lai finished his undergraduate studies in 2008 and his master’s in 2009; his interests crossed computing and cognitive science. Understanding how information is organized was already part of the picture.

Romulus Capital, an early investor, remembered the team as a closely connected group. Its account includes an investor stopping by the Boston office to sign documents just before the founders left for Y Combinator. Crocodoc entered the Winter 2010 batch. The business had a team, a working product and the less photogenic requirement common to young companies: enough money to keep going.

A customer opens another door

Lai’s recollection of those years includes a reduced-gravity NASA project with Damico, well before the Box transaction. Working together came before starting a company together. Later, the founders had to keep working together while their product changed. He remembered an inexperienced team whose persistence proved useful.

“we’re smart and turned out pretty persistent too.”

Peter Lai, in conversation with Scott Orn, 2017

The product’s eventual usefulness did not depend on every customer visiting a Crocodoc homepage. An application programming interface let other businesses incorporate the viewer. Their users could read a document without leaving the place they were already working. That arrangement gave the software a distribution route through someone else’s product, with the viewer doing its job quietly in the background.

Box acquired Crocodoc in May 2013. Later, Box described the technology as the foundation for document viewing in its web and mobile applications, as well as a separate service for developers. The product continued evolving after the acquisition. Box eventually introduced a replacement viewing service and scheduled the older Crocodoc and Box View APIs to retire in January 2018. Useful software has maintenance appointments long after the acquisition photographs.

A budget with a working edge

Lai founded Emburse in 2014. The original business supplied virtual and physical cards for company spending. It belonged to a practical category of financial software: tools that help a manager give someone permission to buy something without granting permission to buy everything.

The distinction is easy to understand at a small company. Someone needs to arrange a trip, pay a vendor or purchase supplies. The organization wants the purchase made, but within limits. Emburse put controls on the payment instrument, allowing managers to set spending parameters and monitor use in real time. A company policy could begin doing work before a transaction became an accounting discussion.

The startup went through Y Combinator’s Winter 2016 batch. YC lists Lai as founder and CEO and Roger Gu as a founder. Its description emphasizes assigning cards to employees or contractors for pre-approved business expenses. The recipient could carry out the task while the company retained a way to define its boundaries. Delegation acquired a few settings.

This is also a change in the employee’s experience. A clear permission is useful to the person doing the buying. It reduces the room for a reasonable purchase to become an unreasonable amount of follow-up. The attraction lies in making a routine decision easier to execute, with less improvisation around who holds the card.

The API makes a second appearance

By October 2016, Emburse was opening card issuance to outside developers. Gu and Lai described customers whose needs were changing as their businesses grew. Early clients might have only 10 or 15 employees. A card product that worked for a small team had to accommodate more elaborate workflows as those teams expanded.

The API let developers incorporate card-generation capabilities into their own tools, including expense applications and chatbots. It was a familiar arrangement from the Crocodoc years: make a useful capability available inside the software where someone already works. A user need not adopt a new destination for every task. Sometimes the new business is the machinery behind the existing destination.

There is a commercial discipline in that choice. A tool has to fit another organization’s habits and constraints. Its success depends partly on work the customer would rather avoid thinking about. The viewer must show the document. The card must respect the limit. Nobody gets extra time in the afternoon because a product has an elegant explanation of itself.

The hotel booking behind the card

Suiteness, which helps travelers book suites and connecting rooms, offers a concrete view of Emburse at work. YC partner Aaron Harris introduced Suiteness founder Kyle Killion to Lai. Killion already knew procurement software from his time at Coupa. The connection had a practical subject: how to handle spending without creating another pile of reconciliation work.

In its June 2019 account, Suiteness described cards expiring before charges arrived and employees having to be tracked down about expenses. It valued Emburse’s instant virtual cards and the ability to manage spending safely. It also singled out the human response when a question arose. There were people at the other end who could help.

The number gives the product a setting beyond a payment screen. Expense administration competes with the work a business exists to do. For a hotel-booking company, time spent hunting for transaction details is time unavailable for booking rooms. The promise of a card with better controls becomes legible in that very ordinary exchange of attention.

The name takes a longer trip

Certify/Chrome River acquired Emburse in July 2019. Financial terms were not disclosed. The card startup entered a portfolio that included Abacus, Captio, Certify, Chrome River, Nexonia and Tallie. The initial plan was to integrate it with Abacus, bringing card controls into a wider expense-management workflow.

Lai described the combination as a way to expand Emburse’s reach while keeping the product usable for frequent and occasional cardholders. That second group matters. Someone who makes a company purchase twice a year has little appetite for becoming an expert in corporate finance software. The tool still has to work when the user has forgotten where the settings live.

In January 2020, the wider group adopted the Emburse name under CEO Eric Friedrichsen. Lai’s original startup and the company subsequently carrying its name belong to different stages of the story. The acquisition brought a card capability into an existing group; the rebranding gave the group a shared name. A short word on a card now covered a much larger collection of software.

Back inside the machinery

By June 2025, Lai was describing a different investigation. He had completed a second iteration of an LLM debugger and was studying how complicated model behavior can arise from simpler components. He credited a year of learning through AI safety programs and mentorship from Stefan Heimersheim. He also announced that he would join Anthropic’s prototyping team the following week.

His public GPT Circuits repository supplies a more concrete picture of the technical work. It contains tools for training toy GPT-2 models and experimenting with sparse autoencoders, with separate configurations, datasets, model definitions and training code. These are deliberately small models. The work gives an interested reader something to inspect rather than requiring confidence in a grand description.

Lai grew up in Los Angeles and returned to California after college. His personal site lists language acquisition, information retrieval and document processing among his interests. It is tempting to arrange such a career as an orderly ascent. The more interesting shape is lateral: a browser document, a controlled purchase, a language model whose internal operations invite examination.

Across those projects, the recurring question seems practical: what is happening inside this thing, and how could someone work with it more easily? That is an interpretation of the work, rather than a slogan Lai has supplied. His record offers plenty to consider without one. There is usually another awkward step between a person and the task they are trying to finish. Lai has repeatedly found something to build there.